Stochastic diffusion model for large-scale temperature downscaling

Generative Deep Learning architectures, such as Diffusion models, offer an alternative to traditional physical modeling and regression models due to their ability to produce stochastic ensembles with a single run. Even though these models are capable of downscaling coarse data, they are often trained in contained regions, which can lead to severe spatial overfitting as the model learns location-specific patterns rather than generalizable physical relationships. In practice, the usability of the models is constrained to the area where they were originally trained, and their predictive capabilities degrade significantly when applied to regions outside the training domain, even if these regions share similar characteristics.
This study presents a one-step and two-step diffusion model capable of downscaling 2-meter temperature from ERA5 to higher-resolution grids in large areas, such as the Contiguous United States or Europe, without spatially overfitting. We use CONUS404, a reanalysis dataset created using simulations of the Weather Research and Forecasting (WRF) model over the Contiguous United States, as our target data and ERA5 and constants involved in the creation of CONUS404, such as altitude and land use, as our input. The model has been trained over the whole area using 10 years of 3-hourly data, and two years have been used for testing. To study the spatial generalization capabilities of the model, we reserve an area of the study region solely for testing and compute evaluation metrics separately for this area to ensure meaningful results. We compare the results of training in large and small areas and the number of years. In addition, we discuss the usefulness of ensemble prediction and the effect that the number of ensemble members has on the performance of the downscaling. Future steps include applying this methodology for downscaling EURO-CORDEX to EMO1 and multivariate downscaling.